HK40079110B - Method, apparatus, device and storage medium for pushing multimedia content - Google Patents

Description

Methods, devices, equipment, and storage media for delivering multimedia content

Technical Field

This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for pushing multimedia content.

Background Technology

Before distributing multimedia content, publishers need to define their target audience as a targeting criterion. Audience targeting is the first step in distributing multimedia content.

In related technologies, for new publishers who need to distribute multimedia content, such as local restaurants or supermarkets, the lack of historical behavioral data means that they can only extract data from people who have visited their city in the past for audience targeting.

Among related technologies, audience targeting has low quality, and multimedia content delivery efficiency is low.

Summary of the Invention

This application provides a method, apparatus, device, and storage medium for pushing multimedia content, which can improve the quality of audience targeting, reduce reliance on historical data, and improve the efficiency of multimedia content push.

According to one aspect of the embodiments of this application, a method for pushing multimedia content is provided, the method comprising:

Obtain service location information and service content information associated with multimedia content;

The service location information and the service content information are input into the joint model for audience targeting to obtain the service geographic features and service preference features of the multimedia content.

The service geographic features and service preference features are respectively fused with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content;

The multimedia content is pushed according to the conversion tendency parameters;

The audience-oriented joint model is trained based on the ranking consistency constraint, which means that the conversion data of a user account for the target multimedia content is positively correlated with the target conversion tendency parameter of the user account for the target multimedia content.

According to one aspect of the embodiments of this application, a multimedia content push device is provided, the device comprising:

The service information acquisition module is used to acquire service location information and service content information associated with multimedia content;

The service feature determination module is used to input the service location information and the service content information into the audience orientation joint model to obtain the service geographic features and service preference features of the multimedia content.

The conversion parameter prediction module is used to fuse the service geographic features and the service preference features with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content.

The content push module is used to push the multimedia content according to the conversion tendency parameter;

The audience-oriented joint model is trained based on the ranking consistency constraint, which means that the conversion data of a user account for the target multimedia content is positively correlated with the target conversion tendency parameter of the user account for the target multimedia content.

According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the above-described method for pushing multimedia content.

According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described method for pushing multimedia content.

According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned multimedia content push method.

The technical solution provided in this application can bring the following beneficial effects:

By setting a ranking consistency constraint that positively correlates user account conversion data and conversion tendency parameters for the same multimedia content, a joint audience targeting model is trained. This allows the model to determine the service geographic characteristics and service preference characteristics of multimedia content using only its service location and content. Combining these service geographic characteristics, service preference characteristics, and account characteristics, the model can predict parameters that measure user conversion rates for multimedia content. Based on these parameters, audience targeting and multimedia content delivery can then be performed. This ranking consistency constraint enables relatively accurate audience targeting for newly released multimedia content even when historical data is scarce or nonexistent. This improves audience targeting quality, reduces reliance on historical data, enhances multimedia content delivery efficiency, avoids wasting computing resources, and alleviates device workload.

Attached Figure Description

To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

Figure 1 is a schematic diagram of the application runtime environment provided in one embodiment of this application;

Figure 2 is a flowchart of a multimedia content push method provided in an embodiment of this application;

Figure 3 is a flowchart of a multimedia content push method provided in an embodiment of this application;

Figure 4 illustrates an exemplary training flowchart for a crowd-oriented joint model;

Figure 5 is a flowchart of a multimedia content push method provided in another embodiment of this application;

Figure 6 illustrates an exemplary process for pushing multimedia content;

Figure 7 is a block diagram of a multimedia content push device provided in an embodiment of this application;

Figure 8 is a structural block diagram of a computer device provided in one embodiment of this application.

Detailed Implementation

The multimedia content push method provided in this application involves artificial intelligence technology and cloud technology, which will be briefly described below to facilitate understanding by those skilled in the art.

Cloud computing refers to the delivery and usage model of IT (Internet Technology) infrastructure, meaning obtaining necessary resources in an on-demand and easily scalable manner through a network. In a broader sense, cloud computing also refers to the delivery and usage model of services, meaning obtaining necessary services in an on-demand and easily scalable manner through a network. These services can be IT and software related, internet-related, or other services. Cloud computing is a product of the development and integration of traditional computer and network technologies such as grid computing, distributed computing, parallel computing, utility computing, network storage technologies, virtualization, and load balancing.

With the development of the internet, real-time data streams, and the diversification of connected devices, as well as the demands for search services, social networks, mobile commerce, and open collaboration, cloud computing has rapidly developed. Unlike previous parallel distributed computing, cloud computing will fundamentally revolutionize the entire internet model and enterprise management model. In the multimedia content push method provided in this application embodiment, multimedia content can be targeted and pushed to specific groups through a cloud server.

Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating/interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning/deep learning, autonomous driving, and intelligent transportation.

Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, smart customer service, vehicle networking, and intelligent transportation. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

The multimedia content push method provided in this application involves the aforementioned artificial intelligence technologies such as natural language processing and machine learning, which will be specifically described through the following embodiments.

To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

Please refer to Figure 1, which shows a schematic diagram of an application runtime environment provided in one embodiment of this application. The application runtime environment may include: terminal 10 and server 20.

Terminal 10 can be an electronic device such as a mobile phone, tablet computer, game console, e-book reader, multimedia playback device, wearable device, or PC (Personal Computer). Application clients can be installed on terminal 10.

In this embodiment, the application can be any application capable of pushing multimedia content. Typically, the application is a news feed content service application. Of course, besides news feed content service applications, other types of applications can also provide multimedia content push services. For example, news applications, social applications, interactive entertainment applications, browser applications, shopping applications, content sharing applications, virtual reality (VR) applications, augmented reality (AR) applications, etc., are not limited in this embodiment. Furthermore, the multimedia content pushed by different applications will vary, and the corresponding functions will also differ. These can be pre-configured according to actual needs, and are not limited in this embodiment. Optionally, the terminal 10 runs a client of the aforementioned application.

Server 20 provides background services to clients of applications in terminal 10. For example, server 20 can be a background server for the aforementioned applications. Server 20 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, server 20 can simultaneously provide background services to applications in multiple terminals 10.

Optionally, terminal 10 and server 20 can communicate with each other via network 30. Terminal 10 and server 20 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

Please refer to Figure 2, which shows a flowchart of a multimedia content push method according to an embodiment of this application. This method can be applied to a computer device, which refers to an electronic device with data computing and processing capabilities. For example, the executing entity for each step can be the server 20 in the application runtime environment shown in Figure 1. The method may include the following steps (210-240).

Step 210: Obtain the service location information and service content information associated with the multimedia content.

The service location associated with the aforementioned multimedia content can also be the geographical location of the multimedia content publisher. This service location information includes the latitude and longitude coordinates of the location. The service content information can be text information within the multimedia content, or it can be tag information obtained by extracting features from the multimedia content. For example, if the multimedia content includes images, semantic recognition can be performed on the images to obtain the corresponding category tags as service content information.

In one possible implementation, the multimedia content is a local life service advertisement. The service location information associated with the multimedia content may be the latitude and longitude data of the service address corresponding to the local life service advertisement. The service content information may be the advertisement copy of the local life service advertisement, or the industry tag corresponding to the local life service advertisement, or a tag added by the advertiser based on the life service content.

Step 220: Input the service location information and service content information into the audience-oriented joint model to obtain the service geographic features and service preference features of multimedia content.

Among them, the audience-oriented joint model is trained based on the ranking consistency constraint, which means that the conversion data of user accounts for target multimedia content is positively correlated with the target conversion tendency parameter of user accounts for target multimedia content.

The aforementioned service geographic features are used to measure the influence of multimedia content on different service areas within a city. The aforementioned service preference features are used to measure the similarity of conversion behaviors among different multimedia content.

The conversion tendency parameters mentioned above can be determined through step 230, that is, the conversion tendency parameters are determined based on service geographic characteristics and service preference characteristics.

For example, if user a converts to multimedia content b more times than user c converts to multimedia content b more times, then user a's conversion tendency parameter value for multimedia content b should also be greater than user c's conversion tendency parameter value for multimedia content b.

In one possible implementation, the aforementioned service geographic features are vector representations of multimedia content in the target geographic feature space, i.e., geographic feature vectors of multimedia content.

In one possible implementation, the aforementioned service preference features are vector representations of multimedia content in the target preference feature space, i.e., the preference feature vectors of multimedia content.

In an exemplary embodiment, as shown in FIG3, FIG3 illustrates a flowchart of a multimedia content push method provided in an embodiment of this application. The above step 220 includes the following sub-steps (221-223).

Step 221: Compare the service location information with the regional location information of each service area to obtain the service geographic features.

The aforementioned regional location information includes the regional center location, such as the latitude and longitude coordinates corresponding to the regional center. By comparing the service location with the regional center location, the geographic feature vector of the multimedia content is obtained.

In one possible implementation, the regional center location of each service area is obtained; based on the service location and the regional center location of each service area, a geographic feature vector of the multimedia content is determined, the geographic feature vector including the probability that a user moves from the k-th regional center to the service location.

Optionally, the method for determining the geographic feature vector of the multimedia content can refer to the method for determining the geographic feature vector _gj of the multimedia content in the embodiments below, which will not be repeated here.

Step 222: Update at least one multimedia content semantic graph based on service location information and service content information to obtain the updated multimedia content semantic graph.

At least one multimedia content semantic graph is used to characterize the degree of association between each multimedia content and different sets of semantic nodes.

Optionally, at least one multimedia content semantic graph includes at least one of the following in the embodiments: content and content semantic graph _GLL , content and conversion time semantic graph _GLT , content and tag semantic graph _GLW , content and region semantic graph _GLR , and content and neighboring user semantic graph _GLU .

Accordingly, the aforementioned set of semantic nodes can be the set of all multimedia content L, the set of conversion time T, the set of all semantic tags W, the set of regions R, and the set of user accounts within 200 meters of the service location of the specified multimedia content during the target time period that have conversion behavior towards the multimedia content.

In one possible implementation, the aforementioned multimedia content is newly released multimedia content. The service location information and service content information of the multimedia content can be added to the multimedia content semantic graph to update the graph structure data of the multimedia content semantic graph.

Step 223: Based on the graph structure data corresponding to the updated multimedia content semantic graph, determine the service preference features of the multimedia content.

In one possible implementation, a graph representation learning method is used to extract features from the graph structure data corresponding to the semantic graph of the multimedia content to determine the preference feature vector of the multimedia content. Optionally, the preference feature vector of the multimedia content may be _vj as described in the embodiments below.

In an exemplary embodiment, as shown in FIG4, FIG4 illustrates a flowchart of the training of a crowd-oriented joint model. The training process of the above-described crowd-oriented joint model includes the following steps (410-430).

Step 410: Obtain sample data logs.

The sample data log includes behavioral data records of sample user accounts for sample multimedia content. The behavioral data records include tag information of the sample multimedia content, and the tag information is the conversion data of sample user accounts for the sample multimedia content.

In some embodiments, the sample data log includes a first log record and a second log record. The recording format of the first log record may be (sample user account, sample multimedia content identifier, number of impressions, number of clicks, number of conversions). The recording format of the second log record may be (sample user account, sample multimedia content identifier, conversion time).

Step 420: Based on the sample data log, determine the service information of the sample multimedia content, the user profile data of the sample user account, and at least one semantic graph of the multimedia content sample.

In one possible implementation, the service information of the aforementioned sample multimedia content can be directly extracted from the sample data log, the user profile data of the aforementioned sample user accounts can be determined by the historical conversion data of the sample user accounts, and the construction method of the aforementioned multimedia content sample semantic graph can refer to the construction methods of content and content semantic graph _GLL , content and conversion time semantic graph _GLT , content and tag semantic graph _GLW , content and region semantic graph _GLR , and content and neighboring user semantic graph _GLU .

Step 430: Based on the service information of the sample multimedia content, the user profile data of the sample user accounts, and at least one semantic graph of the multimedia content samples, train the crowd-oriented joint model according to the joint constraints until the output of the crowd-oriented joint model satisfies the joint constraints.

The joint constraints include order consistency constraints and semantic graph loss conditions.

In one possible implementation, the above joint constraint can be expressed by the mathematical expression (18) in the following embodiment, wherein the above joint constraint is the final loss function satisfying the loss condition.

In one possible implementation, the above-mentioned order consistency constraint can be expressed by the mathematical expression (12) in the following embodiment.

In one possible implementation, the semantic graph loss condition described above can be expressed by the mathematical expression (17) in the following embodiment.

Step 230: The service geographic features and service preference features are fused with the account features of each user account to obtain the conversion tendency parameters of each user account for multimedia content.

In one possible implementation, the conversion tendency parameter is a single value that characterizes the probability of each user account engaging in conversion behavior on multimedia content.

The above fusion can be the calculation of feature distance or the inner product of vectors before feature vectors; this application does not limit this.

In an exemplary embodiment, as shown in FIG3, the above method further includes the following steps (250-270).

Step 250: Obtain the data log.

The data logs include records of user account behavior data for various multimedia content.

In some embodiments, the data logs mentioned above include a first log record and a second log record. Both the first and second log records are behavioral data records of each user account for each multimedia content, and the recorded data may differ. The recording format of the first log record may be (user account, multimedia content identifier, number of impressions, number of clicks, number of conversions). The recording format of the second log record may be (user account, multimedia content identifier, conversion time).

Step 260: Generate at least one multimedia content semantic graph based on the behavioral data records.

Among them, at least one multimedia content semantic graph includes at least one of the following: content and content semantic graph, content and conversion time semantic graph, content and tag semantic graph, content and region semantic graph, and content and neighboring user semantic graph.

Optionally, the specific implementation process of step 260 above can be found in the following methods for constructing the content and content semantic graph _GLL , the content and conversion time semantic graph _GLT , the content and tag semantic graph _GLW , the content and region semantic graph _GLR , and the content and neighboring user semantic graph _GLU .

Step 270: Based on data logs and at least one multimedia content semantic graph, determine the account characteristics of each user account.

In an exemplary embodiment, as shown in FIG5, FIG5 is a flowchart of a multimedia content push method provided in an embodiment of the present application. The aforementioned account features include account geographic features and account preference features, and the aforementioned step 270 includes the following sub-steps (271-272).

Step 271: Input the user profile data of each user account into the audience-oriented joint model for feature extraction processing to obtain the account preference features of each user account.

The user profile data described above is generated based on data logs. The account preference features described above are used to measure the similarity of conversion behavior between different user accounts. Optionally, the account preference features described above can be the user's preference feature vector _vi in the embodiments below.

In one possible implementation, user profile data for each user account is input into a pre-trained neural network model for feature extraction to obtain the user's preference feature vector. The pre-trained neural network model can be composed of a 3-layer neural network, and the parameters of the neural network model are training parameters learned by the model. Furthermore, the neural network model is subject to the aforementioned joint constraint conditions.

Step 272: Input the graph structure data corresponding to at least one multimedia content semantic graph into the audience-oriented joint model to obtain the account geographic features of each user account.

The aforementioned account geographic features are used to measure the influence of a user account across different service areas within a city. In one possible implementation, the aforementioned account geographic features can be a geographic feature vector of the user account. Optionally, the aforementioned account geographic features can be the user's geographic feature vector _gj , including the expected number of conversions of multimedia content by the user account in each service area.

Accordingly, as shown in Figure 5, step 230 above includes the following sub-steps (231-233).

Step 231: For each user account, determine the geographic feature parameters based on the service geographic features and the user account's account geographic features.

In one possible implementation, the aforementioned geographic feature parameters can be the dot product of the geographic feature vector of the multimedia content and the geographic feature vector of the user account, obtained by performing a dot product between the geographic feature vector of the multimedia content and the geographic feature vector of the user account.

Step 232: Determine the preference feature parameters based on the service preference features and the user account's account preference features.

In one possible implementation, the aforementioned preference feature parameters can be the dot product of the preference feature vector of multimedia content and the preference feature vector of user account, obtained by performing a dot product between the preference feature vector of multimedia content and the preference feature vector of user account.

Step 233: Determine the conversion tendency parameters of user accounts for multimedia content based on geographic feature parameters and preference feature parameters.

In one possible implementation, the conversion tendency parameter is the sum of the dot product of the geographic feature vector of the multimedia content and the geographic feature vector of the user account, and the dot product of the preference feature vector of the multimedia content and the preference feature vector of the user account, i.e., the sum of the two dot products.

Optionally, the conversion propensity parameter can be the conversion propensity score y_{_ij} in the examples below.

Finally, conversion preference parameters for multimedia content are generated for each user account.

Step 240: Push multimedia content based on conversion tendency parameters.

In an exemplary embodiment, as shown in FIG3, step 240 above includes the following sub-steps (241-242).

Step 241: Determine the target user account corresponding to the target conversion tendency parameter that meets the conversion tendency parameter conditions.

In one possible implementation, the conversion tendency parameter condition is a threshold condition, such as the conversion tendency parameter value being greater than a preset parameter threshold. Correspondingly, the target conversion tendency parameter is a conversion tendency parameter whose value is greater than the preset parameter threshold.

The target user accounts mentioned above are the user accounts corresponding to the target conversion tendency parameter. The users corresponding to these target user accounts can constitute a user group, which is the result of audience targeting.

Step 242: Push multimedia content to the target user account.

When a target user account is connected to a target traffic domain, multimedia content is pushed to that target user account.

The aforementioned target traffic domain can be the traffic domain for delivering multimedia content, such as the traffic domain of a social application.

In an exemplary embodiment, as shown in FIG5, the above method further includes the following steps (280-310).

Step 280: Obtain the conversion user account corresponding to the multimedia content.

In some embodiments, after new multimedia content is delivered or pushed, conversion behavior may occur, thus allowing the acquisition of the corresponding conversion user accounts. The aforementioned conversion user accounts refer to user accounts that have engaged in conversion behavior related to the multimedia content.

Step 290: Average the account characteristics of the converted user accounts to obtain the average account characteristics of the converted user accounts.

In one possible implementation, the user account's preference feature vector is transformed and subjected to average pooling to obtain the user's average preference feature vector.

Step 300: Based on average account characteristics and service preference characteristics, determine the real-time service preference characteristics of multimedia content.

In one possible implementation, the average preference feature vector is added to the preference feature vector of the multimedia content to obtain the real-time preference feature vector of the multimedia content.

Step 310: Update the conversion tendency parameters based on real-time service preference characteristics.

In one possible implementation, the conversion tendency parameters are recalculated based on the real-time preference feature vector of multimedia content to update the user group corresponding to the target user account, thereby improving the quality of audience targeting and achieving better multimedia content delivery results.

In summary, the technical solution provided in this application, by setting a ranking consistency constraint that shows a positive correlation between user account conversion data and conversion tendency parameters for the same multimedia content, trains a joint audience targeting model. This allows the joint audience targeting model to determine the service geographic characteristics and service preference characteristics of the multimedia content using only its service location and service content. Furthermore, by combining these service geographic characteristics, service preference characteristics, and account characteristics, it can predict parameters that measure user conversion rates for multimedia content. Based on these parameters, audience targeting and multimedia content delivery can be performed. Through the aforementioned ranking consistency constraint, even when there is no historical data or the historical data is sparse, relatively accurate audience targeting can still be achieved for newly released multimedia content. This improves the quality of audience targeting, reduces reliance on historical data, thereby increasing the efficiency of multimedia content delivery, avoiding waste of computing resources, and reducing device operating pressure.

In one example, as shown in Figure 6, a schematic diagram of a multimedia content delivery process is illustrated. The offline process (steps 1 to 4) generates feature vectors related to users and multimedia content daily and writes them into the data engine for rapid retrieval. The online process (step 5) is responsible for: 1) generating audience targeting when creating a local life service advertisement for the first time during multimedia content publishing; and 2) generating new audience targeting conditions for the multimedia content publisher currently running the campaign during automatic updates.

Step 1: Obtain the data logs of the multimedia content.

In one possible implementation, the aforementioned data log includes interaction data between user accounts and historical multimedia content. This interaction data includes exposure data, click data, and conversion data for the historical multimedia content. The exposure data may be the number of user accounts able to view the multimedia content. The click data may be the number of times the multimedia content was interacted with by the user, such as clicking or dragging. The conversion data may be data generated by user conversion behaviors, such as purchasing, making phone calls, inquiring, downloading, or submitting forms after clicking on an advertisement.

In one possible implementation, data logs of historical multimedia content within a target time period are extracted. Optionally, the target time period is the most recent month.

The data log includes a first log record and a second log record. The format of the first log record can be (user account, multimedia content identifier, number of impressions, number of clicks, number of conversions). The format of the second log record can be (user account, multimedia content identifier, conversion time).

In some embodiments, the aforementioned multimedia content is a local life service advertisement. Local life service advertisers, i.e., local life service providers, can publish local life service advertisements in advertising application. In this case, records related to local life service advertisers from the past month can be extracted from the logs of the advertising application, and the record data can be aggregated using the user account (user ID) and advertisement identifier (ad ID) as key fields to obtain a first log record. The record format of the first log record can be (user ID, ad ID, number of impressions, number of clicks, number of conversions). The number of impressions, number of clicks, and number of conversions can be used as the value fields corresponding to the key fields.

Additionally, a historical conversion record (i.e., a second log record) can be extracted. The format of the conversion record can be (User ID, Ad ID, Conversion Time).

Step 2: Generate the account geographic characteristics of the user account and the service geographic characteristics of the multimedia content.

The aforementioned geographic features are used to measure the spatial proximity of user accounts and multimedia content. The geographic features of a user account can be a geographic feature vector of the user account, and the geographic features of multimedia content can be a geographic feature vector of the multimedia content.

For a multimedia content _lj , its geographic feature vector _gj is a K-dimensional vector. Optionally, the mathematical expression (1) for _gj is as follows:

g _j =[f(d(ω _j ,ω ₁ )),...,f(d(ω _j ,ω _K ))] ^T (1)

Where K represents the number of regions, for example, a city can be divided into K regions, i.e., service areas. In some embodiments, the default value of K is 50; _ωj represents the service location of multimedia content _lj , for example, _ωj is a 2D coordinate composed of the longitude and latitude of the location; _ω1 represents the geometric center coordinates of the first region, and the geometric center coordinates of the service location coordinates of all multimedia content in the Kth region are _ωK .

In one possible implementation, given a set of all multimedia content launched in a city in the past year, the K-means algorithm is used to cluster the multimedia content using the service location coordinates of the multimedia content as the clustering feature, resulting in K subsets of multimedia content. Each subset is defined as a region, and the geometric center coordinates of the service location coordinates of all multimedia content in the Kth region are _ωK .

d(*), d(*), and d(*) represent Euclidean distance, such as the Euclidean distance between the service location coordinates of the multimedia content and the geometric center coordinates of the region. The k-th value of the vector _gj represents the probability that the user moves from the center of the k-th region to the service location coordinates of the multimedia content _lj , and the probability value is calculated by the function f(*). The above probability can be the probability of the user accessing the multimedia content, used to characterize the likelihood of the user accessing the multimedia content.

The above f(*) can be any function that produces probability values. In one possible implementation, a Pareto distribution is used to model the relationship between access probability and coordinate distance. Specifically, the mathematical expression (2) for the geographic feature vector _gj of the media content is:

g _j = [(1+d(ω _j , ω ₁ )) ^-α ,..., (1+d(ω _j , ω _K )) ^-α ] ^T (2)

Here, α is the shape parameter of the Pareto distribution, which can be obtained by utilizing the user's conversion sequence across different multimedia content and based on the Maximum Likelihood Estimate (MLE) method.

For the geographic feature vector _{g_i} of user _{u_i} , the mathematical expression (3) of _{g_i} is:

g _i =[γ _i,1 ,γ _i,2 ,...,γ _i,K ] ^T (3)

Where γi _,K represents the user's expected conversion number for multimedia content in the Kth region, and the geographic feature vector corresponding to the user can be obtained by model training.

Step 3: Create a semantic graph of multimedia content.

Constructing a multimedia content semantic graph: To enhance the expressive power of multimedia content in the feature space and accurately generate corresponding preference feature vectors for newly created multimedia content, it is necessary to construct a multimedia content semantic graph to represent the semantic similarity of multimedia content. Then, the semantic similarity is used to generate corresponding preference feature vectors for newly created multimedia content. Each multimedia content semantic graph is a weighted bipartite graph.

In one possible implementation, five types of multimedia content semantic graphs need to be constructed: content-content semantic graph, content-conversion time semantic graph, content-tag semantic graph, content-region semantic graph, and content-nearby user semantic graph.

1. Constructing the content and content semantic graph. In one possible implementation, the mathematical expression (4) of the above content and content semantic graph _GLL is as follows:

G _LL = (L, L, E _LL , W _LL ) (4)

The two sets of nodes on both sides of this bipartite graph are both L, where L represents the set of all multimedia content. The set of edges is E _LL . Using the conversion flow records extracted above, if a user makes conversions on two different multimedia content items within a week, an edge is created between these two multimedia content nodes. The weight of the edge is the number of times the same user accessed these two multimedia content items consecutively within a week, and the set of edge weights is W _LL . The content and content semantic graph is used to represent the sequential access relationships between multimedia content items.

2. Construct a content and conversion time semantic graph. In one possible implementation, the mathematical expression (5) of the above content and conversion time semantic graph _GLT is as follows:

G _LT = (L, T, E _LT , W _LT ) (5)

This bipartite graph has a set of multimedia content L on one side and a set of conversion times T accurate to the hour on the other. Optionally, T contains 24 nodes, each representing one of the 24 hours. The set of edges E_{_LT} represents the number of conversion records for a multimedia content within a certain hour, and the weight of each edge W _{_LT} represents the number of conversions of a multimedia content within that hour.

3. Construct a content and tag semantic graph. In one possible implementation, the mathematical expression (6) of the above content and tag semantic graph _GLW is as follows:

G _LW = (L, W, E _Lw , W _LW ) (6)

This bipartite graph has a multimedia content set L on one side and a semantic tag set W on the other. Optionally, a semantic tag can be a word or phrase contained in the service text copywriting associated with the multimedia content, or it can be the primary and secondary industry tags of the publishing account of the multimedia content. For example, the primary industry tag could be "local life services," and the corresponding secondary industry tags could be subcategories under the primary industry, such as "restaurant," "hairdressing," and "pets." The edge weight represents the frequency of word or phrase occurrences.

4. Construct a content and region semantic graph. In one possible implementation, the mathematical expression (7) of the above content and region semantic graph _GLR is as follows:

G _LR = (L, R, E _LR , W _LR ) (7)

This bipartite graph has a set of multimedia content L on one side and a set of regions R on the other side. E _LR is the set of edges, where each edge represents the location information of a multimedia content belonging to a certain region node. W _LR is the set of weights, where each weight is 1.

5. Construct a content and neighboring user semantic graph. In one possible implementation, the mathematical expression (8) of the above content and neighboring user semantic graph _GLU is as follows:

_GLU = (L, U, _ELU , _WLU ) (8)

This bipartite graph has a set of multimedia content L on one side and a set of user accounts U within a 200-meter radius of the specified multimedia content service location during a target time period (e.g., the past month) that have converted to the multimedia content. The edge weight W _LU represents the number of conversions.

Step 4: Train the crowd-oriented joint model

First, it is necessary to model users and multimedia content in the feature space. This scheme uses vectors to represent the preference feature vectors of user u _i .

In one possible implementation, user profiles stored in the application can be used as user features. The user features are then input into a pre-trained neural network model for feature extraction to obtain the user's preference feature vector. The pre-trained neural network model can be composed of a 3-layer neural network, and the parameters of the neural network model are training parameters learned by the model.

The preference feature vector v_{_j} for multimedia content is obtained entirely by the model. Then, the user's conversion tendency score for multimedia content is defined as y _{_ij} , and the mathematical expression (9) for the conversion tendency score y_{_ij} is as follows:

y _ij ＝v _i ·v _j +g _i ·g _j (9)

This computational method can simultaneously combine the geographic feature vectors and preference feature vectors corresponding to users and multimedia content, and automatically learn the importance of the similarity between preference feature space and geographic feature space in determining user conversion behavior.

Ranking consistency constraint: For a multimedia content, users who convert more times have a higher conversion tendency score than users who convert less times, as expressed by formula (10):

Here, Θ represents the set of users u _i who convert multimedia content, and Θ represents the set of users u _i' whose conversion count C _i'j of multimedia content l _j is less than the conversion count C _ij of user u _i . Θ = {v _i , v _j , g _i | u _i ∈ U, l _j ∈ L} represents the set of training parameters.

P((y _{_ij} - _{y_i′j} )＞0|Θ) represents the probability that user u_{_i} has a higher conversion tendency score than user u_{_i′} .

In one possible implementation, the above probability can be calculated using formula (11), which is as follows:

And it is defined as the loss function corresponding to the set of all multimedia content L, and the mathematical expression (12) is as follows:

Where λ||Θ|| ² is the Gaussian prior parameter used for regularization. Because the loss function used in this embodiment is defined based on the comparison of conversion tendency scores among different users, compared with a classification model that can only use converted users as positive samples and unconverted users as negative samples, it can obtain more training samples and alleviate the problem of decreased model prediction ability caused by insufficient training samples.

Semantic graph loss condition in graph representation learning: For ease of expression, this embodiment uses formula (13) to represent a certain multimedia content semantic graph among the various multimedia content semantic graphs constructed in the above process. Optionally, formula (13) is as follows:

G _LS = (L, S, E _LS , W _Ls ) (13)

Then, the KL divergence (Kullback-Leibler Divergence) between the empirical prediction probability and the model prediction probability is minimized using the LINE (Large-scale Information Network Embedding) model.

Specifically, the probability that a semantic node _{s_k} can be represented by multimedia content _{l_j} is optionally defined as follows: The mathematical expression (14) for s_k is as follows:

Where w _{_jk} is the edge weight. Meanwhile, the probability of using the model to represent the preference feature vector using multimedia content and semantic nodes is p(s _{_k} |l_{_j} ). Optionally, the mathematical expression (15) of p(s _{_k} |l_{_j} ) is as follows:

After simplifying the KL divergence formula, we can obtain the loss function representing a semantic graph of multimedia content. Optionally, the above mathematical expression (16) is as follows:

In this embodiment, since five types of multimedia content semantic graphs are used for representation learning, the objective function (i.e., loss function) used in the graph representation learning part is optional, and its mathematical expression (17) is as follows:

It consists of five sets of semantic nodes.

A joint training model for audience targeting is developed based on ranking consistency constraints and semantic graph loss conditions. In graph representation learning, the multimedia content can be either already deployed or newly created. Newly created multimedia content includes textual material, regions, and nearby users, thus allowing for a biased representation in the feature space, i.e., a biased feature vector. Simultaneously, the newly created multimedia content carries service location information, such as latitude and longitude coordinates, enabling the generation of a geographic feature space representation, i.e., a geographic feature vector.

This embodiment enables the loss function of the multimedia content semantic graph and the ranking consistency loss function to share the embedding features of the multimedia content in the preference space, thereby achieving the purpose of joint training. The final loss function is optionally expressed as follows: The mathematical expression (18) is as follows:

Where β is a coefficient.

In one possible implementation, the parameters of the above model can be iteratively optimized using stochastic gradient descent.

Step 5: Model prediction and online recall of multimedia content.

The user and multimedia content preference and geographic feature vectors generated in the above process can be written to the data engine daily. When a multimedia content publisher needs to target specific audiences, they can first retrieve the geographic and preference feature vectors of the corresponding multimedia content from the data engine. Then, using nearest neighbor retrieval technology, they can find users with the highest conversion tendency scores from the data engine to form an audience targeting group of the specified size. Relevant tags are then added to the multimedia content and written to the multimedia content recall system. When a user in the targeted audience accesses Tencent traffic, the recall system can retrieve that the user is matched by relevant tags, and then extract the multimedia content associated with those tags as a candidate set for exposure, completing the recall step.

In some application scenarios, the solution provided in this embodiment can be applied to multimedia content delivery applications, such as advertising delivery applications. Optionally, a multimedia content-specific audience targeting module can be created in the multimedia content delivery application. This module can execute the solution provided in this embodiment. Taking an advertising delivery application as an example, advertisers only need to specify the location information and advertising category information of the local life service advertisement to be promoted, which can be advertising copy or keywords. Then, they can extract the audience targeting for acquiring new customers on the advertising management platform (DMP) for advertising delivery. In some embodiments, the aforementioned advertising management platform can provide advertisers with different audience extraction methods, such as basic attribute intersection, union, and difference combinations or historical advertising behavior audience extraction methods, mainly used for audience targeting. After obtaining the audience targeting in the DMP, advertisers can bind different audience targeting to different advertisements on the delivery end for advertising delivery.

If a multimedia content publisher needs to update the preference feature vector of their multimedia content hourly, they can first use the audience extraction function for multimedia content to extract the converted users within the last 24 hours, then retrieve the preference feature vectors of these users and perform average pooling to obtain an average user preference feature vector. Then, this average user preference vector is added to the multimedia content's preference vector to obtain a real-time multimedia content preference vector. This vector can then be used to retrieve targeted users from the data engine.

In some application scenarios, the solution provided in this embodiment can also be used to provide advertisers of local life services with an automatic update function. Advertisers of local life services who have already launched advertising can use the automatic update function of the advertising application to modify the audience targeting linked to the ads in a timely manner, thereby improving the effectiveness of the campaign.

In some application scenarios, the solution provided in this embodiment is used to target local life service advertisements, thereby obtaining the target user group for the local life service advertisements. Compared with other solutions, the proportion of new advertisers failing to achieve significant advertising volume is reduced by 20%, and the overall service consumption of local life service advertisements is increased by 10%.

The audience targeting joint model proposed in this embodiment combines ranking consistency and graph representation learning methods to provide audience targeting for local service advertisers. Advertisers only need to specify the location information and advertising category information of the local service to be promoted, which can be ad copy or keywords. Then, they can extract audience targeting for acquiring new customers on the ad management platform for ad placement. In this application scenario, the ranking consistency constraint means that the audience that converts from a local service ad is more likely to convert again than the audience that does not convert from the same ad. This ranking consistency constraint can bring more training samples to the model, alleviating the problem of sparse training samples for local service ads. To measure the strength of conversion tendency between users and local service ads, this solution uses geographic feature vectors and preference feature vectors to compress and represent users and local service ads. The geographic feature vector measures the influence of users and local service ads on different service areas within a city. The preference feature vector measures the similarity of conversion behavior between different users and between different local service ads. To learn the preference feature vector of new local life service advertisements, this embodiment proposes to construct different weighted bipartite graphs to represent different semantic information of the advertisements, and then use a joint training method to integrate the semantic information of the advertisements into the preference feature vector representation learning of the advertisements.

Compared to some related audience targeting solutions, this embodiment can simultaneously improve the performance of local life service advertisements for both new and existing advertisers. It can not only improve the performance of local life service advertisers with existing advertising records, but also enable audience targeting for new advertisers' local life service advertisements, thereby increasing the speed at which new advertisers' advertisements can be generated and achieving the goal of rapid growth. This will increase advertisers' spending on the advertising platform.

The following are embodiments of the apparatus of this application, which can be used to execute embodiments of the method of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method of this application.

Please refer to Figure 7, which shows a block diagram of a multimedia content push device according to an embodiment of this application. This device has the function of implementing the above-described multimedia content push method; the function can be implemented in hardware or by hardware executing corresponding software. The device can be a computer device or can be installed within a computer device. The device 700 may include: a service information acquisition module 710, a service feature determination module 720, a conversion parameter prediction module 730, and a content push module 740.

The service information acquisition module 710 is used to acquire service location information and service content information associated with multimedia content.

The service feature determination module 720 is used to input the service location information and the service content information into the population orientation joint model to obtain the service geographic features and service preference features of the multimedia content.

The conversion parameter prediction module 730 is used to fuse the service geographic features and the service preference features with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content.

The content push module 740 is used to push the multimedia content according to the conversion tendency parameter.

The audience-oriented joint model is trained based on the ranking consistency constraint, which means that the conversion data of a user account for the target multimedia content is positively correlated with the target conversion tendency parameter of the user account for the target multimedia content.

In an exemplary embodiment, the service feature determination module 720 includes: a service geographic feature determination unit, a semantic graph update unit, and a service preference feature determination unit.

The service geographic feature determination unit is used to compare the service location information with the regional location information of each service area to obtain the service geographic features.

A semantic graph update unit is used to update at least one multimedia content semantic graph based on the service location information and the service content information to obtain an updated multimedia content semantic graph, wherein the at least one multimedia content semantic graph is used to characterize the degree of association between each multimedia content and different semantic node sets.

The service preference feature determination unit is used to determine the service preference features of the multimedia content based on the graph structure data corresponding to the updated multimedia content semantic graph.

In an exemplary embodiment, the training process of the crowd-oriented joint model includes:

Obtain sample data logs, which include behavioral data records of sample user accounts for sample multimedia content. The behavioral data records include tag information of the sample multimedia content, and the tag information is the conversion data of the sample user account for the sample multimedia content.

Based on the sample data log, determine the service information of the sample multimedia content, the user profile data of the sample user account, and at least one multimedia content sample semantic graph.

Based on the service information of the sample multimedia content, the user profile data of the sample user account, and the semantic graph of the at least one multimedia content sample, the audience-oriented joint model is trained according to the joint constraint conditions until the output of the audience-oriented joint model satisfies the joint constraint conditions.

The joint constraint conditions include the order consistency constraint conditions and the semantic graph loss conditions.

In an exemplary embodiment, the apparatus 700 further includes: a log acquisition module, a semantic graph generation module, and an account feature determination module.

The log acquisition module is used to acquire data logs, which include behavioral data records of each user account for each multimedia content.

A semantic graph generation module is used to generate the at least one multimedia content semantic graph based on the behavioral data records.

The account feature determination module is used to determine the account features of each user account based on the data log and the at least one multimedia content semantic graph.

The at least one multimedia content semantic graph includes at least one of the following: content and content semantic graph, content and conversion time semantic graph, content and tag semantic graph, content and region semantic graph, and content and neighboring user semantic graph.

In an exemplary embodiment, the account features include account geographic features and account preference features, and the account feature determination module includes: an account preference feature determination unit and an account geographic feature determination unit.

The account preference feature determination unit is used to input the user profile data of each user account into the population-oriented joint model for feature extraction processing to obtain the account preference features of each user account. The user profile data is generated based on the data log.

The account geographic feature determination unit is used to input the graph structure data corresponding to the semantic graph of the at least one multimedia content into the population-oriented joint model to obtain the account geographic features of each user account.

In an exemplary embodiment, the conversion parameter prediction module 730 includes: a geographic feature parameter determination unit, a preference feature parameter determination unit, and a conversion tendency parameter determination unit.

The geographic feature parameter determination unit is used to determine geographic feature parameters for each user account based on the service geographic features and the user account's account geographic features.

The preference feature parameter determination unit is used to determine preference feature parameters based on the service preference features and the account preference features of the user account.

The conversion tendency parameter determination unit is used to determine the conversion tendency parameter of the user account for the multimedia content based on the geographic feature parameter and the preference feature parameter.

In an exemplary embodiment, the apparatus 700 further includes: a conversion account acquisition module, an average account characteristic determination module, a service preference characteristic update module, and a conversion tendency parameter update module.

The conversion account acquisition module is used to acquire the conversion user account corresponding to the multimedia content;

The average account feature determination module is used to average the account features of the converted user accounts to obtain the average account features of the converted user accounts.

The service preference feature update module is used to determine the real-time service preference features of the multimedia content based on the average account features and the service preference features.

The conversion tendency parameter update module is used to update the conversion tendency parameter according to the real-time service preference characteristics.

In summary, the technical solution provided in this application, by setting a ranking consistency constraint that shows a positive correlation between user account conversion data and conversion tendency parameters for the same multimedia content, trains a joint audience targeting model. This allows the joint audience targeting model to determine the service geographic characteristics and service preference characteristics of the multimedia content using only its service location and service content. Furthermore, by combining these service geographic characteristics, service preference characteristics, and account characteristics, it can predict parameters that measure user conversion rates for multimedia content. Based on these parameters, audience targeting and multimedia content delivery can be performed. Through the aforementioned ranking consistency constraint, even when there is no historical data or the historical data is sparse, relatively accurate audience targeting can still be achieved for newly released multimedia content. This improves the quality of audience targeting, reduces reliance on historical data, thereby increasing the efficiency of multimedia content delivery, avoiding waste of computing resources, and reducing device operating pressure.

It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

Please refer to Figure 8, which shows a structural block diagram of a computer device provided in one embodiment of this application. This computer device can be a server for executing the aforementioned multimedia content push method. Specifically:

Computer device 800 includes a central processing unit (CPU) 801, a system memory 804 including random access memory (RAM) 802 and read-only memory (ROM) 803, and a system bus 805 connecting the system memory 804 and the CPU 801. Computer device 800 also includes a basic input/output system (I/O system) 806 that facilitates information transfer between various devices within the computer, and a mass storage device 807 for storing the operating system 813, application programs 814, and other program modules 815.

The basic input/output system 806 includes a display 808 for displaying information and an input device 809 for user input, such as a mouse or keyboard. Both the display 808 and the input device 809 are connected to the central processing unit 801 via an input/output controller 810 connected to the system bus 805. The basic input/output system 806 may also include the input/output controller 810 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input/output controller 810 also provides output to a display screen, printer, or other types of output devices.

Mass storage device 807 is connected to central processing unit 801 via a mass storage controller (not shown) connected to system bus 805. Mass storage device 807 and its associated computer-readable media provide non-volatile storage for computer device 800. That is, mass storage device 807 may include computer-readable media (not shown) such as hard disk or CD-ROM (CompactDisc Read-Only Memory) drive.

Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 804 and mass storage device 807 described above can be collectively referred to as memory.

According to various embodiments of this application, the computer device 800 can also be connected to a remote computer on a network, such as the Internet, for operation. That is, the computer device 800 can be connected to a network 812 via a network interface unit 811 connected to the system bus 805, or the network interface unit 811 can be used to connect to other types of networks or remote computer systems (not shown).

The memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the above-described method for pushing multimedia content.

In an exemplary embodiment, a computer-readable storage medium is also provided, the storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set, when executed by a processor, implements the above-described method for pushing multimedia content.

Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned multimedia content delivery method.

It should be understood that "multiple" as used herein refers to two or more. "And/or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and/or B can represent: A alone, A and B simultaneously, or B alone. The character "/" generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims (17)

1. A method for pushing multimedia content, characterized in that the method includes: Obtain service location information and service content information associated with multimedia content; The service location information and the service content information are input into the joint model for audience targeting to obtain the service geographic features and service preference features of the multimedia content. The service geographic features and service preference features are respectively fused with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content; The multimedia content is pushed according to the conversion tendency parameters; The audience targeting joint model is trained based on the ranking consistency constraint, which means that the number of conversions corresponding to the conversion data of a user account for the target multimedia content is positively correlated with the target conversion tendency parameter of the user account for the target multimedia content. 2. The method according to claim 1, characterized in that, inputting the service location information and the service content information into a joint model for audience targeting to obtain the service geographic features and service preference features of the multimedia content includes: The service location information is compared with the regional location information of each service area to obtain the service geographic features; Based on the service location information and the service content information, at least one multimedia content semantic graph is updated to obtain an updated multimedia content semantic graph. The at least one multimedia content semantic graph is used to characterize the degree of association between each multimedia content and different sets of semantic nodes. Based on the graph structure data corresponding to the updated multimedia content semantic graph, the service preference features of the multimedia content are determined. 3. The method according to claim 2, characterized in that the training process of the crowd-oriented joint model includes: Obtain sample data logs, which include behavioral data records of sample user accounts for sample multimedia content. The behavioral data records include tag information of the sample multimedia content, and the tag information is the conversion data of the sample user account for the sample multimedia content. Based on the sample data log, determine the service information of the sample multimedia content, the user profile data of the sample user account, and at least one multimedia content sample semantic graph. Based on the service information of the sample multimedia content, the user profile data of the sample user account, and the semantic graph of the at least one multimedia content sample, the audience-oriented joint model is trained according to the joint constraint conditions until the output of the audience-oriented joint model satisfies the joint constraint conditions. The joint constraint conditions include the order consistency constraint conditions and the semantic graph loss conditions. 4. The method according to claim 2, characterized in that the method further comprises: Obtain data logs, which include behavioral data records of each user account for each multimedia content; Based on the recorded behavioral data, at least one semantic graph of multimedia content is generated; Based on the data logs and the at least one multimedia content semantic graph, the account characteristics of each user account are determined; The at least one multimedia content semantic graph includes at least one of the following: content and content semantic graph, content and conversion time semantic graph, content and tag semantic graph, content and region semantic graph, and content and neighboring user semantic graph. 5. The method according to claim 4, wherein the account features include account geographic features and account preference features, and the step of determining the account features of each user account based on the data log and the at least one multimedia content semantic graph includes: The user profile data of each user account is input into the population-oriented joint model for feature extraction processing to obtain the account preference features of each user account. The user profile data is generated based on the data log. The graph structure data corresponding to the semantic graph of the at least one multimedia content is input into the joint model for audience targeting to obtain the account geographic features of each user account. 6. The method according to claim 5, characterized in that, the step of fusing the service geographic features and the service preference features with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content includes: For each user account, geographical feature parameters are determined based on the service geographical features and the user account's account geographical features; Based on the service preference features and the account preference features of the user account, the preference feature parameters are determined; Based on the geographic feature parameters and the preference feature parameters, the conversion tendency parameters of the user account for the multimedia content are determined. 7. The method according to any one of claims 1 to 6, characterized in that the method further comprises: Obtain the conversion user account corresponding to the multimedia content; The account characteristics of the converted user accounts are averaged to obtain the average account characteristics of the converted user accounts; Based on the average account characteristics and the service preference characteristics, the real-time service preference characteristics of the multimedia content are determined; The conversion tendency parameter is updated based on the real-time service preference characteristics. 8. A multimedia content delivery device, characterized in that the device comprises: The service information acquisition module is used to acquire service location information and service content information associated with multimedia content; The service feature determination module is used to input the service location information and the service content information into the audience orientation joint model to obtain the service geographic features and service preference features of the multimedia content. The conversion parameter prediction module is used to fuse the service geographic features and the service preference features with the account features of each user account to obtain the conversion tendency parameters of each user account for the multimedia content. The content push module is used to push the multimedia content according to the conversion tendency parameter; The audience targeting joint model is trained based on the ranking consistency constraint, which means that the number of conversions corresponding to the conversion data of a user account for the target multimedia content is positively correlated with the target conversion tendency parameter of the user account for the target multimedia content. 9. The apparatus according to claim 8, wherein the service feature determination module comprises: a service geographic feature determination unit, a semantic graph update unit, and a service preference feature determination unit; The service geographic feature determination unit is used to compare the service location information with the regional location information of each service area to obtain the service geographic features; A semantic graph update unit is used to update at least one multimedia content semantic graph based on the service location information and the service content information to obtain an updated multimedia content semantic graph. The at least one multimedia content semantic graph is used to characterize the degree of association between each multimedia content and different sets of semantic nodes. The service preference feature determination unit is used to determine the service preference features of the multimedia content based on the graph structure data corresponding to the updated multimedia content semantic graph. 10. The apparatus according to claim 9, wherein the apparatus further comprises: The sample log acquisition module is used to acquire sample data logs, which include behavioral data records of sample user accounts for sample multimedia content. The behavioral data records include tag information of the sample multimedia content, and the tag information is the conversion data of the sample user account for the sample multimedia content. The sample information determination module is used to determine, based on the sample data log, the service information of the sample multimedia content, the user profile data of the sample user account, and at least one semantic graph of the multimedia content sample. The model training module is used to train the audience-oriented joint model based on the service information of the sample multimedia content, the user profile data of the sample user accounts, and the semantic graph of the at least one multimedia content sample, according to joint constraints, until the output of the audience-oriented joint model satisfies the joint constraints; wherein, the joint constraints include the ranking consistency constraint and the semantic graph loss condition. 11. The apparatus according to claim 9, wherein the apparatus further comprises: a log acquisition module, a semantic graph generation module, and an account feature determination module; The log acquisition module is used to acquire data logs, which include behavioral data records of each user account for each multimedia content; A semantic graph generation module is used to generate at least one semantic graph of multimedia content based on the behavioral data records; The account feature determination module is used to determine the account features of each user account based on the data log and the at least one multimedia content semantic graph; wherein, the at least one multimedia content semantic graph includes at least one of the following: content and content semantic graph, content and conversion time semantic graph, content and tag semantic graph, content and region semantic graph, and content and neighboring user semantic graph. 12. The apparatus according to claim 11, wherein the account features include account geographic features and account preference features, and the account feature determination module includes: an account preference feature determination unit and an account geographic feature determination unit; The account preference feature determination unit is used to input the user profile data of each user account into the population-oriented joint model for feature extraction processing to obtain the account preference features of each user account. The user profile data is generated based on the data log. The account geographic feature determination unit is used to input the graph structure data corresponding to the semantic graph of the at least one multimedia content into the population-oriented joint model to obtain the account geographic features of each user account. 13. The apparatus according to claim 12, wherein the conversion parameter prediction module comprises: a geographic feature parameter determination unit, a preference feature parameter determination unit, and a conversion tendency parameter determination unit; The geographic feature parameter determination unit is used to determine geographic feature parameters for each user account based on the service geographic features and the user account's account geographic features. The preference feature parameter determination unit is used to determine preference feature parameters based on the service preference features and the account preference features of the user account. The conversion tendency parameter determination unit is used to determine the conversion tendency parameter of the user account for the multimedia content based on the geographic feature parameter and the preference feature parameter. 14. The apparatus according to any one of claims 8-13, characterized in that the apparatus further comprises: a conversion account acquisition module, an average account characteristic determination module, a service preference characteristic update module, and a conversion tendency parameter update module; The conversion account acquisition module is used to acquire the conversion user account corresponding to the multimedia content; The average account feature determination module is used to average the account features of the converted user accounts to obtain the average account features of the converted user accounts. The service preference feature update module is used to determine the real-time service preference features of the multimedia content based on the average account features and the service preference features. The conversion tendency parameter update module is used to update the conversion tendency parameter according to the real-time service preference characteristics. 15. A computer device, characterized in that the computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the multimedia content push method as described in any one of claims 1 to 7. 16. A computer-readable storage medium, characterized in that the storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the multimedia content push method as described in any one of claims 1 to 7. 17. A computer program product, comprising a computer program, characterized in that, when the computer program is executed by an electronic device, it implements the method for pushing multimedia content as described in any one of claims 1 to 7.